Re: [Qemu-devel] SeaBIOS, FW_CFG_NUMA, and FW_CFG_MAX_CPUS

[Eduardo Habkost <ehabkost@redhat.com>]

On Wed, Jul 25, 2012 at 05:47:36PM +0300, Gleb Natapov wrote:
> On Fri, Jul 20, 2012 at 05:00:25PM -0300, Eduardo Habkost wrote:
> > Hi,
> > 
> > While working at the CPU index vs APIC ID changes, I stumbled upon
> > another not-very-well-defined interface between SeaBIOS and QEMU, and I
> > would like to clarify the semantics and constraints of some FW_CFG
> > entries.
> > 
> > First, the facts/assumptions:
> > 
> > - There's no concept of "CPU index" or "CPU identifier" that SeaBIOS and
> >   QEMU agree upon, except for the APIC ID. All SeaBIOS can really see
> >   are the CPU APIC IDs, on boot or on CPU hotplug.
> > - The APIC ID is already a perfectly good CPU identifier, that is
> >   present on bare metal hardware too.
> >   - Adding a new kind of "CPU identifier" in addition to the APIC ID
> >     would just make things more complex. 
> >   - The only problem with APIC IDs is that they may not be contiguous.
> > 
> > That said, I would like to clarify the meaning of:
> > 
> > - FW_CFG_MAX_CPUS
> > 
> > What are the basic semantics and expectations about FW_CFG_MAX_CPUS?
> > Considering that the APIC IDs may not be contiguous, is it supposed to
> > be:
> > 
> > a) the maximum number of CPUs that will be ever online, doesn't matter
> >    their APIC IDs, or
> > b) a value so that every CPU has APIC ID < MAX_CPUS.
> > 
> > A practical example: suppose we have a machine with 18 CPUs with the
> > following APIC IDs: 0x00, 0x01, 0x02, 0x04, 0x05, 0x06, 0x08, 0x09,
> > 0x0a, 0x10, 0x11, 0x12, 0x14, 0x15, 0x16, 0x18, 0x19, 0x1a.
> > 
> > (That's the expected result for a machine with 2 sockets, 3 cores per
> > socket, 3 threads per core.)
> > 
> > In that case, should FW_CFG_MAX_CPUS be: a) 18, or b) 27 (0x1b)?
> > 
> > If it should be 18, it will require additional work on SeaBIOS to make:
> > - CPU hotplug work
> > - SRAT/MADT/SSDT tables be built with Processor ID != APIC ID
> > - SRAT/MADT/SSDT tables be kept stable if the system is hibernated and
> >   resumed after a CPU is hot-plugged.
> > 
> > (Probably in that case I would suggest introducing a FW_CFG_MAX_APIC_ID
> > entry, so that SeaBIOS can still build the ACPI tables more easily).
> > 
> > 
> > - FW_CFG_NUMA
> > 
> > The first problem with FW_CFG_NUMA is that it depends on FW_CFG_MAX_CPUS
> > (so it inherits the same questions above). The second is that
> > FW_CFG_NUMA is a CPU-based table, but there's nothing SeaBIOS can use to
> > know what CPUs FW_CFG_NUMA is refering to, except for the APIC IDs. So,
> > should FW_CFG_NUMA be indexed by APIC IDs?
> > 
> > 
> > - My proposal:
> > 
> > My proposal is to try to keep things simple, and just use the following
> > rule:
> > 
> >  - Never have a CPU with APIC ID > FW_CFG_MAX_CPUS.
> > 
> > This way:
> > - The SeaBIOS ACPI code can be kept simple.
> > - The current CPU hotplug interface can work as-is (up to 256 CPUs),
> >   based on APIC IDs.
> > - The current FW_CFG_NUMA interface can work as-is, indexed by APIC IDs.
> > - The ACPI tables can be easily kept stable between hibernate and
> >   resume, after CPU hotplug.
> > 
> > This is the direction I am trying to go, and I am sending this just to
> > make sure nobody is against it, and to not surprise anybody when I send
> > a QEMU patch to make FW_CFG_MAX_CPUS be based on APIC IDs.
> > 
> This shouldn't change the meaning of maxcpus on command line though.
> Qemu can calculate max ACPI ID needed to support maxcpus by itself.

Yes, that's my plan. The command-line will hide the concept of "APIC
IDs", but make sure the the APIC-ID-based QEMU<=>SeaBIOS interface
works.

In other words, FW_CFG_MAX_CPUS would be set to:

 apic_id_for_cpu_index(max_cpus - 1) + 1


> 
> > 
> > My second proposal would be to introduce a FW_CFG_MAX_APIC_ID entry, so
> > the SeaBIOS ACPI code can be kept simple.
> > 
> > My third proposal would be to introduce a FW_CFG CPU Index => APIC ID
> > table, but I really wouldn't like to introduce a new type of CPU
> > identifier to be used between QEMU and SeaBIOS, when the APIC ID is a
> > perfectly good unique CPU identifier that already exists in bare metal
> > hardware.
> > 
> > -- 
> > Eduardo
> 
> --
> 			Gleb.
> 

-- 
Eduardo